Automating Academic Literature Reviews with Doubao Work’s AI Agent

Background

Repeating the same manual steps—searching literature, extracting data, classifying papers, and writing a coherent report—consumes most of a researcher’s time. The author tried to solve this by delegating the whole workflow to Doubao Work’s AI Agent.

Step 1: Literature Search – Batch Extraction and Automatic Classification

The author added a specialized Academic Literature Research skill to the agent. Using the arXiv API (invoked via curl), the agent fetched the latest 30 Agent Harness papers, automatically classified each paper, and produced an Excel sheet with 15 columns (e.g., index, title, authors, author count, arXiv ID, submission date, research‑topic classification, etc.).

From arXiv collect the newest 30 Agent Harness papers, classify them, and output an Excel file with one row per paper and the columns mentioned above.

This step, which previously required a full morning of manual reading and tagging, was completed in a few minutes, leaving only a brief review and minor adjustments.

Step 2: Analysis and Visualization – Let the Data Speak

With the classification table ready, the author asked the agent to analyse trends. The agent used matplotlib to generate:

A bar chart of the number of papers per research category (highlighting "Harness architecture design and vertical‑domain deployment" and "RL‑based harness training").

A line chart of publication volume over time.

A comparative chart of average author count per category.

Based on these visualisations, the agent suggested two concrete research topics:

Agent Harness formal composition and reparability verification (a hot, small‑team‑friendly direction).

Tool‑calling conflict and coordination mechanisms in multi‑agent Harness (an unexplored area).

Step 3: Drafting the Review

Having the tables, charts, and statistical insights, the author instructed the agent to produce a PowerPoint‑style literature review. The requested structure was:

Background and significance

Method categories (one subsection per category)

Comparative analysis of methods

Research trends and hotspots

Future outlook

The agent generated a coherent draft, automatically inserting citations (author and year) and labeling axes and legends in the figures.

Custom Skill “Review Workshop”

The entire three‑step pipeline—literature analysis, data visualisation, and review generation—was saved as a custom Skill named Review Workshop . By invoking this Skill later, the same workflow can be rerun with a single command, without leaving the agent’s conversation window.

Observations

The process worked seamlessly across phone, PC, and web, allowing the author to continue refining requirements on a mobile device while the heavy‑weight tasks ran on a cloud‑based virtual desktop. The author stresses that the agent does not replace creative research, hypothesis formulation, or experimental design, but it reliably handles repetitive, mechanical tasks that drain enthusiasm.

The service offers a 30‑day free trial, enabling users to evaluate the productivity gains without immediate cost.